PGD Attack Step

Problem

One projected gradient descent attack step: x_adv = x + alpha * sign(grad), then project each entry back into the epsilon ball [x - epsilon, x + epsilon].

The signature is pgd_step(x, grad, epsilon, alpha). Both x and grad are 2D matrices.

Hint

PGD is FGSM followed by a projection back into the allowed epsilon ball.

Test cases

case 1

in: [[[1],[1]],[[1],[1]],0.1,0.05]

expected: [[1.05],[1.05]]

case 2

in: [[[0],[1]],[[1],[1]],0.5,1]

expected: [[0.5],[1.5]]

case 3

in: [[[1],[0]],[[0],[0]],0.2,0.1]

expected: [[1],[0]]

case 4

in: [[[0.2],[1]],[[1],[1]],0.5,0.4]

expected: [[0.6],[1.4]]

Starter codePython

def pgd_step(x, grad, epsilon, alpha):
    # Your code here
    pass

4 test cases · run your code against them in the editor